Optimizing CPU Scheduling and Inter-VM Communication for Lightweight Virtualization using Reinforcement Learning

doi:10.1016/j.future.2026.108612

Optimizing CPU Scheduling and Inter-VM Communication for Lightweight Virtualization using Reinforcement Learning

2026-06-09

0

PRE

AI

DOI：10.1016/j.future.2026.108612

Original

Original request for help

Save

Abstract

En 中文

In recent years, lightweight virtualization has gained significant attention for consolidating Mixed-Criticality Systems (MCS) in Internet of Things (IoT) and edge computing deployments. With growing demands for computational efficiency and real-time responsiveness, optimizing CPU scheduling and inter-virtual-machine (VM) communication has become crucial. Conventional schedulers often fail to efficiently balance dynamic workloads, resulting in suboptimal resource utilization and elevated latency. To overcome these limitations, this paper introduces a comprehensive optimization framework for CPU scheduling in virtualized environments. Our approach enhances the native Priority-Based Monotonic (PRM) scheduler and incorporates a contextual bandit (CB) reinforcement learning (RL) algorithm for intelligent vCPU resource management, substantially boosting real-time performance. We also implement custom CPU and memory utilization monitoring instructions to enable efficient system introspection. The entire framework is ported and validated on the emerging PhytiumPi platform. Experimental evaluation demonstrates that our enhanced Xvisor (eXvisor) hypervisor successfully boots multiple guest operating systems on PhytiumPi with a minimal overhead of approximately 1.05% as measured by CoreMark. The CB-based scheduler not only achieves a remarkable 96.3% latency reduction in real-time tasks compared to the native Priority-Based Round Robin (PRR) algorithm but also directly translates these gains into superior communication performance. Specifically, it boosts Remote Processor Messaging (RPMsg) throughput by 48% (to 4.74 Mbps) and reduces virtual network latency by 14% (to 0.392 ms) over baseline schedulers. These results validate that our framework significantly improves both real-time performance and communication efficiency, offering a practical and highly effective solution for resource-constrained MCS in IoT and edge environments.

Optimizing CPU Scheduling and Inter-VM Communication for Lightweight Virtualization using Reinforcement Learning

Abstract

Journal

Organization

Cited Papers